Tiny Aya goes local-first for multilingual AI

Hot take: this is the right shape of multilingual AI for the real world. Instead of chasing ever-larger general models, Cohere is optimizing for locality, language depth, and deployment constraints that actually matter outside well-connected markets.

• –The regional split is the main differentiator: Earth, Fire, and Water suggest a more disciplined approach than a single generic multilingual checkpoint.
• –3.35B parameters is small enough to be operationally useful on constrained hardware, which broadens adoption beyond enterprise cloud setups.
• –The underserved-language focus is strategically stronger than broad-but-shallow coverage, especially for education and public-interest deployments.
• –This reads as both a research release and a practical product play: useful for developers, but also aligned with sovereignty, access, and offline inference.